from __future__ import annotations import json import math from pathlib import Path from typing import Any import numpy as np import torch import torchvision from einops import rearrange from PIL import Image from torch.utils.data import Dataset import os from datasets import load_dataset, DownloadConfig from tqdm import tqdm from PIL import ImageFile ImageFile.LOAD_TRUNCATED_IMAGES = True class FinetuneDataset(Dataset): def __init__( self, path: str = '', split: str = "train", splits: tuple[float, float, float] = (0.9, 0.05, 0.05), min_resize_res: int = 256, max_resize_res: int = 256, crop_res: int = 256, flip_prob: float = 0.5, msr_vtt_cc_full: bool = False, mix: list[str] = ['magicbrush', 'something', 'hq'], mix_factors: list[float] = [40, 1, 1], copy_prob: float = 0.0, kubric_100k: bool = False, ): self.split = split assert split in ("train", "val", "test") assert sum(splits) == 1 self.path = path self.min_resize_res = min_resize_res self.max_resize_res = max_resize_res self.crop_res = crop_res self.flip_prob = flip_prob self.mix_factors = mix_factors self.msr_vtt_cc_full = msr_vtt_cc_full self.copy_prob = copy_prob self.data = [] for dataset in mix: if dataset != 'hq': for _ in range(mix_factors[mix.index(dataset)]): if kubric_100k and dataset == 'kubric': self.data.extend(json.load(open(f'data/{dataset}/train_100k.json', 'r'))) print("LODADED KUBRIC 100K") else: self.data.extend(json.load(open(f'data/{dataset}/train.json', 'r'))) # if dataset == 'msr-vtt-cc': # self.data.extend(json.load(open(f'data/{dataset}/train_gpt.json', 'r'))) if split == 'val': self.data = self.data[:2] def __len__(self) -> int: return len(self.data) def __getitem__(self, i: int) -> dict[str, Any]: # if i < len(self.data): ex = self.data[i] img_path0 = ex['input'] img_path1 = ex['output'] prompt = ex['instruction'] dataset = img_path0.split('/')[1] if dataset == 'kubric': subtask = img_path0.split('/')[2] else: subtask = '___' if type(prompt) == list: prompt = prompt[0] spatial = 'left' in prompt.lower() or 'right' in prompt.lower() image_1 = Image.open(img_path1).convert('RGB') if i < len(self.data) else img_path1 if subtask not in ['closer', 'counting', 'further_location', 'rotate']: if self.copy_prob > 0 and torch.rand(1) < self.copy_prob: image_0 = Image.open(img_path1).convert('RGB') if i < len(self.data) else img_path1 else: image_0 = Image.open(img_path0).convert('RGB') if i < len(self.data) else img_path0 else: image_0 = Image.open(img_path0).convert('RGB') if i < len(self.data) else img_path0 reize_res = torch.randint(self.min_resize_res, self.max_resize_res + 1, ()).item() image_0 = image_0.resize((reize_res, reize_res), Image.Resampling.LANCZOS) image_1 = image_1.resize((reize_res, reize_res), Image.Resampling.LANCZOS) image_0 = rearrange(2 * torch.tensor(np.array(image_0)).float() / 255 - 1, "h w c -> c h w") image_1 = rearrange(2 * torch.tensor(np.array(image_1)).float() / 255 - 1, "h w c -> c h w") crop = torchvision.transforms.RandomCrop(self.crop_res) flip_prob = 0.0 if spatial else self.flip_prob flip = torchvision.transforms.RandomHorizontalFlip(float(flip_prob)) image_0, image_1 = flip(crop(torch.cat((image_0, image_1)))).chunk(2) return dict(edited=image_1, edit=dict(c_concat=image_0, c_crossattn=prompt)) class MagicEditDataset(Dataset): def __init__( self, path: str = '../../change_descriptions/something-something', split: str = "train", splits: tuple[float, float, float] = (0.9, 0.05, 0.05), min_resize_res: int = 256, max_resize_res: int = 256, crop_res: int = 256, flip_prob: float = 0.0, debug: bool = False, ): self.min_resize_res = min_resize_res self.max_resize_res = max_resize_res self.crop_res = crop_res self.flip_prob = flip_prob print("Dataset params") print(self.min_resize_res, self.max_resize_res, self.crop_res, self.flip_prob) #clean json (if first and last are not both present, remove) split = "train" if split == "train" else "dev" self.dataset = load_dataset("osunlp/MagicBrush")[split] # if split == 'dev': # eval_data = json.load(open('eval_data/video_edit.json', 'r')) # dummy_image = Image.new('RGB', (1, 1), (0, 0, 0)) # eval_data = { # 'source_img': [Image.open(x['img0']) for x in eval_data], # 'target_img': [Image.open(x['img1']) for x in eval_data], # 'instruction': [x['edit'] if type(x['edit']) == str else x['edit'][0] for x in eval_data], # 'img_id': ['' for _ in eval_data], # 'turn_index': np.array([1 for _ in eval_data], dtype=np.int32), # 'mask_img': [dummy_image for _ in eval_data] # Replace each entry with the dummy image # } # eval_dataset = HuggingFaceDataset.from_dict(eval_data) # self.dataset = concatenate_datasets([self.dataset, eval_dataset]) if debug: self.dataset = self.dataset.shuffle(seed=42).select(range(50)) def __len__(self) -> int: return len(self.dataset) def __getitem__(self, i: int) -> dict[str, Any]: prompt = self.dataset[i]['instruction'] if type(prompt) == list: prompt = prompt[0] image_0 = self.dataset[i]['source_img'] image_1 = self.dataset[i]['target_img'] if image_0.mode == 'RGBA': image_0 = image_0.convert('RGB') if image_1.mode == 'RGBA': image_1 = image_1.convert('RGB') reize_res = torch.randint(self.min_resize_res, self.max_resize_res + 1, ()).item() image_0 = image_0.resize((reize_res, reize_res), Image.Resampling.LANCZOS) image_1 = image_1.resize((reize_res, reize_res), Image.Resampling.LANCZOS) image_0 = rearrange(2 * torch.tensor(np.array(image_0)).float() / 255 - 1, "h w c -> c h w") image_1 = rearrange(2 * torch.tensor(np.array(image_1)).float() / 255 - 1, "h w c -> c h w") crop = torchvision.transforms.RandomCrop(self.crop_res) flip = torchvision.transforms.RandomHorizontalFlip(float(self.flip_prob)) image_0, image_1 = flip(crop(torch.cat((image_0, image_1)))).chunk(2) return dict(edited=image_1, edit=dict(c_concat=image_0, c_crossattn=prompt)) class FrameEditDataset(Dataset): def __init__( self, path: str = '../../change_descriptions/something-something', split: str = "train", splits: tuple[float, float, float] = (0.9, 0.05, 0.05), task: str = 'flickr30k_text', min_resize_res: int = 256, max_resize_res: int = 256, crop_res: int = 256, flip_prob: float = 0.0, debug: bool = False, ): self.split = split self.task = task if split == "train": path = os.path.join(path, 'train.json') self.json = json.load(open(path, 'r')) np.random.shuffle(self.json) self.min_resize_res = min_resize_res self.max_resize_res = max_resize_res self.crop_res = crop_res self.flip_prob = flip_prob #clean json (if first and last are not both present, remove) if split == 'train': new_json = [] for i in range(len(self.json)): video_id = self.json[i]['id'] img_path0 = f'../../change_descriptions/something-something/frames/{video_id}/first.jpg' img_path1 = f'../../change_descriptions/something-something/frames/{video_id}/last.jpg' if os.path.exists(img_path0) and os.path.exists(img_path1): new_json.append(self.json[i]) self.json = new_json if debug: self.json = self.json[:50] def __len__(self) -> int: return len(self.json) def __getitem__(self, i: int) -> dict[str, Any]: if self.split == 'train': video_id = self.json[i]['id'] img_path0 = f'../../change_descriptions/something-something/frames/{video_id}/first.jpg' img_path1 = f'../../change_descriptions/something-something/frames/{video_id}/last.jpg' prompt = self.json[i]['label'] image_0 = Image.open(img_path0).convert('RGB') image_1 = Image.open(img_path1).convert('RGB') reize_res = torch.randint(self.min_resize_res, self.max_resize_res + 1, ()).item() # image_0 = image_0.resize((reize_res, reize_res), Image.Resampling.LANCZOS) # image_1 = image_1.resize((reize_res, reize_res), Image.Resampling.LANCZOS) image_0 = image_0.resize((self.crop_res, self.crop_res)) image_1 = image_1.resize((self.crop_res, self.crop_res)) image_0 = rearrange(2 * torch.tensor(np.array(image_0)).float() / 255 - 1, "h w c -> c h w") image_1 = rearrange(2 * torch.tensor(np.array(image_1)).float() / 255 - 1, "h w c -> c h w") crop = torchvision.transforms.RandomCrop(self.crop_res) flip = torchvision.transforms.RandomHorizontalFlip(float(self.flip_prob)) image_0, image_1 = flip(crop(torch.cat((image_0, image_1)))).chunk(2) # if i ever wanna reverse time # if torch.rand(1) > 0.5: # image_0, image_1 = image_1, image_0 # prompt = caption0 if self.split == 'train': return dict(edited=image_1, edit=dict(c_concat=image_0, c_crossattn=prompt)) else: return dict(edited=image_1, edit=dict(c_concat=image_0, c_crossattn=texts)) class EditITMDataset(Dataset): def __init__( self, path: str = '../../change_descriptions/something-something', split: str = "test", splits: tuple[float, float, float] = (0.9, 0.05, 0.05), task: str = 'flickr30k_text', min_resize_res: int = 256, max_resize_res: int = 256, crop_res: int = 256, flip_prob: float = 0.0, debug: bool = False, ): self.split = split self.task = task # if task == 'flickr_edit': # path = 'data/flickr_edit/valid.json' if split == 'val' else 'data/flickr_edit/test.json' # self.json = json.load(open(path, 'r')) # #clean json, if "pos" key is empty string, remove # self.json = [x for x in self.json if x['pos'] != ''] if task == 'whatsup': path = 'data/whatsup/itm_test.json' if split == 'test' else 'data/whatsup/itm_valid.json' self.json = json.load(open(path, 'r')) elif task == 'svo': path = 'data/svo/itm_test.json' if split == 'test' else 'data/svo/itm_valid.json' self.json = json.load(open(path, 'r')) else: path = f'data/{task}/valid.json' self.json = json.load(open(path, 'r')) self.json = [x for x in self.json if x.get('pos', '') != ''] self.min_resize_res = min_resize_res self.max_resize_res = max_resize_res self.crop_res = crop_res self.flip_prob = flip_prob if debug: self.json = self.json[:50] def __len__(self) -> int: return len(self.json) def __getitem__(self, i: int) -> dict[str, Any]: ex = self.json[i] pos = ex.get('pos', '') if pos == '': pos = ex['prompt'] neg = ex.get('neg', '') if neg == '': neg = ex['prompt'] img_path0 = ex['input'] texts = [pos, neg] # if self.task == 'whatsup' or self.task == 'svo': # img_path0 = f"data/{self.task}/images/{ex['image']}" if self.task == 'flickr_edit' else ex['image'] # texts = [ex['pos'], ex['neg']] # else: # img_path0 = ex['input'] # texts = ex['pos'], ex['prompt'] # subtasks = ex['type'] if self.task == 'flickr_edit' else '' try: image_0 = Image.open(img_path0).convert('RGB') reize_res = torch.randint(self.min_resize_res, self.max_resize_res + 1, ()).item() image_0 = image_0.resize((reize_res, reize_res), Image.Resampling.LANCZOS) image_0 = rearrange(2 * torch.tensor(np.array(image_0)).float() / 255 - 1, "h w c -> c h w") except: image_0 = 0 return dict(input=image_0, texts=texts, path=img_path0) class OldFrameEditDataset(Dataset): def __init__( self, path: str = '../../change_descriptions/something-something', split: str = "train", splits: tuple[float, float, float] = (0.9, 0.05, 0.05), task: str = 'flickr30k_text', min_resize_res: int = 256, max_resize_res: int = 256, crop_res: int = 256, flip_prob: float = 0.0, debug: bool = False, ): if split == "train": path = os.path.join(path, 'train.json') elif split == "val": path = os.path.join(path, 'validation.json') self.json = json.load(open(path, 'r')) np.random.shuffle(self.json) self.min_resize_res = min_resize_res self.max_resize_res = max_resize_res self.crop_res = crop_res self.flip_prob = flip_prob #clean json (if first and last are not both present, remove) new_json = [] for i in range(len(self.json)): video_id = self.json[i]['id'] img_path0 = f'../../change_descriptions/something-something/frames/{video_id}/first.jpg' img_path1 = f'../../change_descriptions/something-something/frames/{video_id}/last.jpg' if os.path.exists(img_path0) and os.path.exists(img_path1): new_json.append(self.json[i]) self.json = new_json if debug: self.json = self.json[:50] def __len__(self) -> int: return len(self.json) def __getitem__(self, i: int) -> dict[str, Any]: video_id = self.json[i]['id'] img_path0 = f'../../change_descriptions/something-something/frames/{video_id}/first.jpg' img_path1 = f'../../change_descriptions/something-something/frames/{video_id}/last.jpg' prompt = self.json[i]['label'] image_0 = Image.open(img_path0) image_1 = Image.open(img_path1) reize_res = torch.randint(self.min_resize_res, self.max_resize_res + 1, ()).item() image_0 = image_0.resize((reize_res, reize_res), Image.Resampling.LANCZOS) image_1 = image_1.resize((reize_res, reize_res), Image.Resampling.LANCZOS) image_0 = rearrange(2 * torch.tensor(np.array(image_0)).float() / 255 - 1, "h w c -> c h w") image_1 = rearrange(2 * torch.tensor(np.array(image_1)).float() / 255 - 1, "h w c -> c h w") crop = torchvision.transforms.RandomCrop(self.crop_res) flip = torchvision.transforms.RandomHorizontalFlip(float(self.flip_prob)) image_0, image_1 = flip(crop(torch.cat((image_0, image_1)))).chunk(2) # if i ever wanna reverse time # if torch.rand(1) > 0.5: # image_0, image_1 = image_1, image_0 # prompt = caption0 return dict(edited=image_1, edit=dict(c_concat=image_0, c_crossattn=prompt)) class EditDataset(Dataset): def __init__( self, path: str = 'data/clip-filtered-dataset', split: str = "train", splits: tuple[float, float, float] = (0.9, 0.05, 0.05), min_resize_res: int = 256, max_resize_res: int = 256, crop_res: int = 256, flip_prob: float = 0.0, ): self.split = split assert split in ("train", "val", "test") assert sum(splits) == 1 self.path = path self.min_resize_res = min_resize_res self.max_resize_res = max_resize_res self.crop_res = crop_res self.flip_prob = flip_prob self.genhowto = open('data/genhowto/genhowto_train_clip0.7_filtered.txt', 'r').readlines() # self.genhowto = open('data/genhowto/genhowto_train.txt', 'r').readlines() self.genhowto = [x.strip() for x in self.genhowto] new_genhowto = [] for i in range(len(self.genhowto)): img_path, prompt, prompt2 = self.genhowto[i].split(':') new_genhowto.append((img_path, prompt, 'action')) new_genhowto.append((img_path, prompt2, 'state')) self.genhowto = new_genhowto with open(Path(self.path, "seeds.json")) as f: self.seeds = json.load(f) split_0, split_1 = { "train": (0.0, splits[0]), "val": (splits[0], splits[0] + splits[1]), "test": (splits[0] + splits[1], 1.0), }[split] idx_0 = math.floor(split_0 * len(self.seeds)) idx_1 = math.floor(split_1 * len(self.seeds)) self.seeds = self.seeds[idx_0:idx_1] # shuffle seeds and genhowto # np.random.seed(42) # np.random.shuffle(self.seeds) # np.random.shuffle(self.genhowto) def __len__(self) -> int: return len(self.seeds) + len(self.genhowto) def __getitem__(self, i: int) -> dict[str, Any]: if i < len(self.seeds): name, seeds = self.seeds[i] propt_dir = Path(self.path, name) seed = seeds[torch.randint(0, len(seeds), ()).item()] with open(propt_dir.joinpath("prompt.json")) as fp: prompt = json.load(fp)["edit"] image_0 = Image.open(propt_dir.joinpath(f"{seed}_0.jpg")) image_1 = Image.open(propt_dir.joinpath(f"{seed}_1.jpg")) else: ex = self.genhowto[i - len(self.seeds)] # img_path, prompt, prompt2 = ex.split(':') # img_path = img_path.replace('changeit_detected_without_test', 'changeit_detected') # img_path = 'data/genhowto/' + img_path # full_img = Image.open(img_path).convert('RGB') # image_0 = full_img.crop((0, 0, full_img.width // 3, full_img.height)) # image_1 = full_img.crop((full_img.width * 2 // 3, 0, full_img.width, full_img.height)) # image_2 = full_img.crop((full_img.width // 3, 0, full_img.width * 2 // 3, full_img.height)) # if torch.rand(1) > 0.5: # image_1 = image_2 # prompt = prompt2 img_path, prompt, type = ex img_path = img_path.replace('changeit_detected_without_test', 'changeit_detected') img_path = 'data/genhowto/' + img_path full_img = Image.open(img_path).convert('RGB') image_0 = full_img.crop((0, 0, full_img.width // 3, full_img.height)) if type == 'action': image_1 = full_img.crop((full_img.width // 3, 0, full_img.width * 2 // 3, full_img.height)) else: image_1 = full_img.crop((full_img.width * 2 // 3, 0, full_img.width, full_img.height)) reize_res = torch.randint(self.min_resize_res, self.max_resize_res + 1, ()).item() image_0 = image_0.resize((reize_res, reize_res), Image.Resampling.LANCZOS) image_1 = image_1.resize((reize_res, reize_res), Image.Resampling.LANCZOS) image_0 = rearrange(2 * torch.tensor(np.array(image_0)).float() / 255 - 1, "h w c -> c h w") image_1 = rearrange(2 * torch.tensor(np.array(image_1)).float() / 255 - 1, "h w c -> c h w") crop = torchvision.transforms.RandomCrop(self.crop_res) flip = torchvision.transforms.RandomHorizontalFlip(float(self.flip_prob)) image_0, image_1 = flip(crop(torch.cat((image_0, image_1)))).chunk(2) return dict(edited=image_1, edit=dict(c_concat=image_0, c_crossattn=prompt)) class GenHowToDataset(Dataset): def __init__( self, path: str = 'data/clip-filtered-dataset', split: str = "train", splits: tuple[float, float, float] = (0.9, 0.05, 0.05), min_resize_res: int = 256, max_resize_res: int = 256, crop_res: int = 256, flip_prob: float = 0.0, ): self.split = split assert split in ("train", "val", "test") assert sum(splits) == 1 self.path = path self.min_resize_res = min_resize_res self.max_resize_res = max_resize_res self.crop_res = crop_res self.flip_prob = flip_prob self.genhowto = open('data/genhowto/genhowto_train.txt', 'r').readlines() self.genhowto = [x.strip() for x in self.genhowto] new_genhowto = [] for i in range(len(self.genhowto)): img_path, prompt, prompt2 = self.genhowto[i].split(':') new_genhowto.append((img_path, prompt, 'action')) new_genhowto.append((img_path, prompt2, 'state')) self.genhowto = new_genhowto np.random.shuffle(self.genhowto) # with open(Path(self.path, "seeds.json")) as f: # self.seeds = json.load(f) # split_0, split_1 = { # "train": (0.0, splits[0]), # "val": (splits[0], splits[0] + splits[1]), # "test": (splits[0] + splits[1], 1.0), # }[split] # idx_0 = math.floor(split_0 * len(self.seeds)) # idx_1 = math.floor(split_1 * len(self.seeds)) # self.seeds = self.seeds[idx_0:idx_1] # shuffle seeds and genhowto # np.random.seed(42) # np.random.shuffle(self.seeds) # np.random.shuffle(self.genhowto) def __len__(self) -> int: return len(self.genhowto) def __getitem__(self, i: int) -> dict[str, Any]: ex = self.genhowto[i] img_path, prompt, type = ex img_path = img_path.replace('changeit_detected_without_test', 'changeit_detected') img_path = 'data/genhowto/' + img_path full_img = Image.open(img_path).convert('RGB') image_0 = full_img.crop((0, 0, full_img.width // 3, full_img.height)) if type == 'action': image_1 = full_img.crop((full_img.width // 3, 0, full_img.width * 2 // 3, full_img.height)) else: image_1 = full_img.crop((full_img.width * 2 // 3, 0, full_img.width, full_img.height)) reize_res = torch.randint(self.min_resize_res, self.max_resize_res + 1, ()).item() image_0 = image_0.resize((reize_res, reize_res), Image.Resampling.LANCZOS) image_1 = image_1.resize((reize_res, reize_res), Image.Resampling.LANCZOS) image_0 = rearrange(2 * torch.tensor(np.array(image_0)).float() / 255 - 1, "h w c -> c h w") image_1 = rearrange(2 * torch.tensor(np.array(image_1)).float() / 255 - 1, "h w c -> c h w") crop = torchvision.transforms.RandomCrop(self.crop_res) flip = torchvision.transforms.RandomHorizontalFlip(float(self.flip_prob)) image_0, image_1 = flip(crop(torch.cat((image_0, image_1)))).chunk(2) return dict(edited=image_1, edit=dict(c_concat=image_0, c_crossattn=prompt)) class EditDatasetEval(Dataset): def __init__( self, path: str, split: str = "train", splits: tuple[float, float, float] = (0.9, 0.05, 0.05), res: int = 256, ): assert split in ("train", "val", "test") assert sum(splits) == 1 self.path = path self.res = res with open(Path(self.path, "seeds.json")) as f: self.seeds = json.load(f) split_0, split_1 = { "train": (0.0, splits[0]), "val": (splits[0], splits[0] + splits[1]), "test": (splits[0] + splits[1], 1.0), }[split] idx_0 = math.floor(split_0 * len(self.seeds)) idx_1 = math.floor(split_1 * len(self.seeds)) self.seeds = self.seeds[idx_0:idx_1] def __len__(self) -> int: return len(self.seeds) def __getitem__(self, i: int) -> dict[str, Any]: name, seeds = self.seeds[i] propt_dir = Path(self.path, name) seed = seeds[torch.randint(0, len(seeds), ()).item()] with open(propt_dir.joinpath("prompt.json")) as fp: prompt = json.load(fp) edit = prompt["edit"] input_prompt = prompt["input"] output_prompt = prompt["output"] image_0 = Image.open(propt_dir.joinpath(f"{seed}_0.jpg")) reize_res = torch.randint(self.res, self.res + 1, ()).item() image_0 = image_0.resize((reize_res, reize_res), Image.Resampling.LANCZOS) image_0 = rearrange(2 * torch.tensor(np.array(image_0)).float() / 255 - 1, "h w c -> c h w") return dict(image_0=image_0, input_prompt=input_prompt, edit=edit, output_prompt=output_prompt)